4. Parallel Computing 14. Introduction

Networking and Client-Server Parallelism

Another type of distributed memory computer is a computer network. In this type of

system, several computers, each capable of operating stand-alone, are interconnected.

The interconnection scheme is typically in the nature of a bus, such as an Ethernet®.

These networks were not originally conceived for purposes of parallel computing as

much as they were for information sharing. However, they can be used for parallel

computing effectively if the degree of locality is very high.

A common paradigm for parallel computing in a network is known as client-server. In

this model, long-term processes running on selected nodes of the network provide a

service for processes on other nodes. The latter, called clients, communicate with the

server by sending it messages. Many servers and clients can be running concurrently. A

given node might support both client and server processes. Furthermore, a server of one

function might be a client of others. The idea is similar to object-oriented computing,

except that the objects in this case run concurrently with each other, rather than being

dormant until the next message is received.

From the programmer's point of view, communication between client and server takes

place using data abstractions such as sockets. The socket concept permits establishment

of a connection between a client and a server by the client knowing the server's address in

the network. Knowing the address allows the client to send the server an initial connect

message. After connection is established, messages can be exchanged without the need,

by the program, to use the address explicitly. Of course, the address is still used implicitly

to route the message from one node to another. A common form of socket is called the

stream socket. In this form, once the connection is established, reading and writing

appears to be an ordinary i/o operation, similar to using streams in C++.

---

606

Parallel Computing